96 lines
4.2 KiB
Markdown
96 lines
4.2 KiB
Markdown
# Model Recommendations — Best Price/Performance to Add to Admin-AI
|
||
|
||
**Research date:** July 9, 2026
|
||
**Authored by:** Research Team via Sho'Nuff
|
||
|
||
---
|
||
|
||
## Already Covered (Direct Through Admin-AI — No OpenRouter)
|
||
|
||
| Model | In/Out per 1M | Class |
|
||
|---|---|---|
|
||
| deepseek-chat (V3) | $0.14 / $0.28 | Budget generalist |
|
||
| deepseek-v4-pro | $0.435 / $0.87 | Strong reasoning |
|
||
| deepseek-v4-flash | $0.15 / $0.30 | Fast budget |
|
||
| deepseek-reasoner (R1) | $0.55 / $2.19 | Deep reasoning |
|
||
| deepseek-coder | $0.14 / $0.28 | Code specialist |
|
||
| gemini-flash-latest | $0.08–0.30 / $0.30–2.50 | Cheapest |
|
||
| gemini-pro-latest | $1.25 / $5.00 | Strong vision |
|
||
| claude-haiku-4-5 | $1.00 / $5.00 | Fast Anthropic |
|
||
| claude-sonnet-4-6 | $3.00 / $15.00 | Balanced Anthropic |
|
||
| claude-opus-4-8 | $5.00 / $25.00 | Best reasoning |
|
||
| GPT-5.5 | $5.00 / $30.00 | OpenAI flagship |
|
||
| grok-4.3 | ~$3.00+ | xAI reasoning |
|
||
|
||
## 🔥 Top Additions — Best Price/Performance
|
||
|
||
### Tier 1: Must Add (Cheap + Strong)
|
||
|
||
| Model | In/Out per 1M | Context | Provider | Why Add |
|
||
|---|---|---|---|---|
|
||
| **Qwen 3.5 27B** | ~$0.35 / $0.35 | 128K | Together / Groq / Fireworks | Best open-weight coding model. SWE-bench 77% — beats Claude Sonnet on code at 1/40th the price |
|
||
| **Qwen 3.6 Plus** | ~$0.80 / $0.80 | 128K | Alibaba Cloud | Stronger than 3.5, 120B class reasoning |
|
||
| **Mistral Small** | $0.10 / $0.30 | 128K | Mistral | European, GDPR-friendly. Cheapest general-purpose option |
|
||
| **Llama 4 (70B)** | ~$0.60 / $0.90 | 1M | Together / Groq | Open-weight. 1M context window at $0.60 input. Good generalist |
|
||
| **GPT-5.5 Mini** | $1.50 / $6.00 | 256K | OpenAI | Budget GPT — strong reasoning at 1/3 flagship price |
|
||
|
||
### Tier 2: Strong Value (Specific Use Cases)
|
||
|
||
| Model | In/Out per 1M | Context | Provider | Why Add |
|
||
|---|---|---|---|---|
|
||
| **Gemini 3.1 Flash-Lite** | $0.10 / $0.40 | 1M | Google | Even cheaper than Flash. 1M context. For bulk tasks |
|
||
| **Mistral Large** | $2.00 / $6.00 | 128K | Mistral | European Claude competitor. GDPR-compliant. Strong reasoning |
|
||
| **Grok 4 Mini** | $0.30 / $0.50 | 128K | xAI | xAI's budget model. Already have Grok 4.3, add the cheap one |
|
||
| **Llama 4 405B** | $3.50 / $3.50 | 128K | Together | Largest open model. For tasks needing maximum open-model power |
|
||
|
||
### Tier 3: Specialty Models
|
||
|
||
| Model | In/Out per 1M | Context | Provider | Why Add |
|
||
|---|---|---|---|---|
|
||
| **Qwen 3.6 Coder** | ~$0.35 / $0.35 | 128K | Alibaba | Code-focused version of Qwen's best |
|
||
| **Codestral** | ~$0.30 / $0.90 | 256K | Mistral | Mistral's code-specialized model |
|
||
| **MiniMax M2.5** | ~$0.50 / $2.00 | 256K | MiniMax | Chinese budget option, 1M context |
|
||
|
||
---
|
||
|
||
## Cost Comparison — Current vs Recommended
|
||
|
||
| Scenario | All DeepSeek (current) | With Qwen/Mistral mixed | Annual savings |
|
||
|---|---|---|---|
|
||
| 1M requests/month (avg 1000 in, 500 out tokens) | ~$360/mo | ~$90/mo | **$3,240/yr** |
|
||
| 10M requests/month | ~$3,600/mo | ~$900/mo | **$32,400/yr** |
|
||
|
||
Qwen 3.5 27B at $0.35/$0.35 handles 80% of coding tasks at 1/3 the cost of deepseek-v4-pro. Mistral Small at $0.10/$0.30 handles simple Q&A at less than half of deepseek-chat.
|
||
|
||
---
|
||
|
||
## How to Add Them
|
||
|
||
All of these are available through API gateways already supported by LiteLLM (admin-ai):
|
||
|
||
**Via Together AI** (single API key, multiple models):
|
||
- Qwen 3.5 27B: `together_ai/qwen/Qwen3.5-27B`
|
||
- Llama 4 70B: `together_ai/meta-llama/Llama-4-70B`
|
||
- Qwen 3.6 Coder: `together_ai/qwen/Qwen3.6-Coder`
|
||
|
||
**Via Groq** (fastest inference, free tier available):
|
||
- Qwen 3.5: `groq/qwen-3.5-27b`
|
||
- Llama 4: `groq/llama-4-70b`
|
||
- Mistral models: via Groq or direct
|
||
|
||
**Via Mistral API** (direct):
|
||
- Mistral Small: `mistral/mistral-small-latest`
|
||
- Mistral Large: `mistral/mistral-large-latest`
|
||
|
||
**Via xAI API** (likely already configured):
|
||
- Grok 4 Mini: `xai/grok-4-mini`
|
||
|
||
---
|
||
|
||
## Immediate Recommendation
|
||
|
||
1. **Get a Together AI API key** ($5 credit free) — gives Qwen 3.5 27B + Llama 4 + Mistral access through one account
|
||
2. **Add `together_ai/qwen/Qwen3.5-27B`** — single best price/performance model available anywhere right now
|
||
3. **Add `mistral/mistral-small-latest`** — $0.10/$0.30 is unbeatable for simple tasks
|
||
4. **Add the Together models to admin-ai's LiteLLM config** — one API key unlocks 5+ high-value models
|